Gestalt Labs

Independent Canadian post-training research lab (founded 2025, founder Dan Lougen / DJLougen). Thesis: naive abliteration — subtracting a refusal direction from weights — destroys capability-entangled representations (8+ point TruthfulQA drops, high KL from base). Their alternative, with full pipeline transparency (SFT, RL, configs, training code all open):

• SABER (Spectral Analysis-Based Entanglement Resolution) — canonical correlation analysis identifies which latent directions encode refusal vs. useful capability, then surgically edits only the refusal-correlated subspace. Measured: 92.54% compliance proxy, HarmBench ASR 0.67%, KLD 0.0083, PPL ratio 1.10.
• NSC-ACE (Neural Steering Committee for Agentic Co-Evolution) — GRPO-based RL that trains steering vectors alongside policy optimization; better instruction-following and latent controllability.

On the doombox: Ornstein3.6-27B-MTP-NSC-ACE-SABER (Q4_K_M-MTP, 9 of 59 llama-swap entries) — a Qwen3.6-27B dense base through the full pipeline (NSC-ACE → SABER → Ornstein SFT refinement, rank 32/32 LoRA-stage for reasoning/personality, 1 epoch) with a verified MTP head (llama.cpp PR #22673 conversion). The other locally-downloaded GestaltLabs dirs (Qwen3.6-35B-A3B-NSC-ACE-SABER ±MTP) are the MoE flagship line, not registered in llama-swap.

Also releases: Ornstein-Hermes (Hermes-tuned + SABER), BOREAL (from-scratch pretraining), BusyBeaver (tool-policy safety research for agentic AI — small models learning to follow tool-use policies and refuse unsafe tool calls).

Links: Qwen model family (their bases) · LoRA/SFT/LLM fine-tuning (the techniques) · OrcaRouter (the naive-abliteration counterpoint they define themselves against)